<p>Fine-grained identification of volcanic lithologies in heterogeneous reservoirs is highly challenging. The Huoshiling Formation in the Wangfu Fault Depression (Songliao Basin) contains 18 lithology classes with high inter-class similarity, severe class imbalance, and cross-well distribution shifts. In this study, we jointly use thin-section images and conventional well logs to systematically evaluate two representative multimodal fusion strategies: a feature-level Cross-Attention Fusion Network (CAFN) that maximizes cross-modal interaction, and a decision-level Late Fusion Network (LFN) that fuses calibrated probabilities from unimodal experts. In addition, we include unimodal expert models, an early-concatenation MLP (EarlyConcat-MLP), a lightweight Transformer-based fusion model (TCMT), and a CCA-based joint embedding model (CCA-Embed) as additional baselines. All models share the same training configuration and are compared under a strictly grouped, leakage-free 5-fold protocol. Results indicate that decision-level fusion is more favorable than further increasing feature-level interaction complexity in this setting. LFN achieves the best overall trade-off among accuracy, macro-averaged metrics, probability calibration, and computational cost: it obtains 99.24% overall accuracy (98.76% macro-F1) under grouped 5-fold evaluation and 97.74% overall accuracy (97.3% macro-F1) in cross-regional blind-well validation. LFN also exhibits higher stability in robustness experiments with injected noise and reduced training data. Furthermore, an explainable AI framework combining Grad-CAM and SHAP shows that LFN relies more on geologically plausible cues, providing interpretive support for its stronger cross-well generalization. Overall, this work supports a “robustness over complexity” perspective for multimodal fusion in this scenario and offers a transparent, reusable framework for evaluating geoscience AI models.</p>